AMPREDICT Mobility-4: A Novel Four Category Mobility Outcome Prediction Model for Patients Undergoing Vascular Amputation

Abstract

Objective:

The aim of this study was to create a new mobility outcome prediction model (AMPREDICT Mobility-4) that can be applied at the time a decision is made that amputation is necessary due to an underlying vascular aetiology. It can be used to predict mobility outcome in patients undergoing transmetatarsal (TM), transtibial (TT), or transfemoral (TF) amputation.

Methods:

A cohort study retrospectively identified persons with lower limb amputation (LLA) through a large Veterans Affairs dataset, then prospectively enrolled participants to obtain their 12 month post-amputation self reported mobility. Seven hundred and four patients with first unilateral TM, TT, or TF amputation secondary to diabetes and or peripheral arterial disease were identified between February 2021 and September 2022. Potential predictors incorporated factors from the following domains: prior revascularisation; amputation level; demographics; comorbidities; mental health; and health behaviour factors. The predicted mobility outcome included four functional levels: wheelchair mobility; household ambulatory mobility; basic community ambulation; and advanced community ambulation. Multinomial logistic regression was used to fit 1 year post-incident amputation risk prediction models. Variable selection was performed using lasso, a machine learning methodology.

Results:

Variable selection led to a final model of 15 predictors that successfully discriminated between all categories of mobility, except household vs. basic community mobility. The model discriminated best when comparing wheelchair mobility with at least some community ambulatory mobility (optimism adjusted c index = 0.71), a basic level of community mobility vs. less than a basic level of community mobility (optimism adjusted c index = 0.70), and advanced community mobility vs. less than advanced community mobility (optimism adjusted c index = 0.73).

Conclusion:

The AMPREDICT Mobility-4 model predicted four categories of functional mobility at each of three LLA levels using predictors available in the electronic health record. The model can augment clinical personalised mobility outcome prediction, enhancing both surgeon and patient awareness of the consequences of amputation level on mobility, to facilitate patient surgeon amputation level shared decision-making and assisting in setting appropriate patient outcome expectations.

INTRODUCTION

Learning of the need for lower limb amputation (LLA) secondary to complications of diabetes and or peripheral arterial disease (PAD) results in significant psychological distress.¹ The uncertainty of future mobility and its effect on function is a major concern.^2,3 There is a complex interplay of trade offs between amputation level and important patient outcomes, especially in the individual patient. Often, more proximal amputations are associated with a reduced risk of failed healing; more distal amputations are associated with greater mobility, but these potential mobility benefits may not be realised because of the increased risk of delayed healing, the need for ongoing wound care, and ultimately re-amputation at a higher level,⁴ which can negatively impact ambulation.^5–8 Realisation of higher levels of mobility is often associated with reduced mortality risks.^9–11 While mobility is only one of these amputation level trade offs, it is the single most important determinant of quality of life in a person undergoing LLA.³

Decision support tools developed using prediction models are increasingly being used to assist providers in complex clinical decisions. Such tools allow clinicians to consider and discuss patients’ personalised outcome risks, rather than average risks, especially in populations that have very heterogeneous outcomes. Few prediction models have been developed to inform mobility outcomes in the peri-amputation period^12–14 and only one, AMPREDICT-Mobility, was developed for use in the pre-surgical period.¹³

AMPREDICT-Mobility predicts two categories of mobility: independent basic and advanced mobility.¹³ The model includes several predictors available in the electronic health record (EHR); however, four predictors require patient interview, thus increasing clinical burden. Furthermore, it lacks the granularity to predict more functionally nuanced categories of functional mobility, such as wheelchair, household, and different levels of community mobility.

The aim of this study was to develop a new prediction model (AMPREDICT Mobility-4) that predicts the probability of achieving four distinct categories of mobility 1 year after amputation in individuals undergoing their initial LLA secondary to vascular disease or diabetes at either transmetatarsal (TM), transtibial (TT), or transfemoral (TF) amputation levels, using only predictors readily available in the EHR. The overarching goal was to provide higher resolution of mobility outcomes that are important in treatment planning, setting patient expectations, and ensuring a low clinical burden.

MATERIALS AND METHODS

Study design

This cohort study retrospectively identified persons with LLA through a large Veterans Affairs (VA) dataset, then prospectively collected their self reported mobility (Supplementary Table S1).

Data source

The VA Corporate Data Warehouse (CDW) houses data from multiple VA clinical and administrative systems. The CDW was accessed to identify patients aged ≥ 30 years undergoing their first diabetes and or PAD related amputation at the TM, TT, or TF level between February 2021 and September 2022, as determined by Current Procedural Terminology (CPT) and International Classification of Diseases, 9th and 10th revisions (ICD-9 and ICD-10) procedure codes. Amputations at digit or ray level were excluded, as were patients with a prior amputation at TM level or higher, those undergoing a bilateral amputation or an amputation due to trauma, or if they had a diagnosis of paraplegia, quadriplegia, spinal cord injury, dementia, or a body mass index (BMI) < 15 kg/m² or > 52 kg/m². All codes and first amputations and laterality have been published previously.¹⁵ Participants were eligible for participation if they were not excluded based on the above criteria and they did not die in the 12 months post-amputation (and thus had a 1 year mobility outcome).

Patient recruitment

A modification of the recruitment method successfully used in previous studies was employed.¹⁶ This process was performed in monthly waves. Patients were contacted in chronological order to obtain their outcome at 12 months (range, 10 – 14 months) post-amputation via a mailed introductory letter, the Amputee Single Item Mobility Measure (AMPSIMM) questionnaire, and an opt out postcard; if they did not return a completed questionnaire or opt out postcard within two weeks, potential participants were contacted via telephone and were interviewed by a study co-ordinator. If potential participants did not return the first questionnaire, did not opt out, and were unreachable by telephone, a final questionnaire was mailed. Patient baseline characteristics and candidate predictors among those eligible who did and did not consent to provide their mobility were compared to rule out potential bias in the outcome assessment. All procedures were approved by the local institutional review board.

Candidate predictor variables

The prediction model evaluated factors available within the EHR that fell into several key domains: amputation level; demographics; comorbidities; mental health factors; health behaviours; and prior revascularisation (Table 1).^13,17–19 The EHR, an electronic version of the patient medical record, is becoming universal across most health systems. All predictors were identified through the CDW using corresponding ICD and CPT codes where appropriate. Race and sex were not evaluated as candidate predictors due to concerns about their use in prediction models. Their association with outcome may be related to surrogate factors that could not be incorporated into model development (e.g., socioeconomic status, access to care, mental health, etc.) and which may provide predictions that could lead to disparate care.^20–22 All 32 candidate predictors preceded the amputation, considered time zero for the prediction model.

Table 1.

Description of all candidate predictors evaluated for the AMPREDICT Mobility-4 prediction model.

Predictor*	Description, when applicable
Amputation level	Transmetatarsal, transtibial, transfemoral
Demographics †
Age	At time of incident amputation
Marital status	Nominal (married, single, or/never married, separated or divorced, vs. widowed); binary (married vs. not married)
Urban vs. rural or highly rural	Current living environment
BMI	In kg/m2, measured prior to amputation
Comorbidities ‡
Asthma
Kidney dialysis
Coronary atherosclerosis
Diabetes	Any diabetes (diabetes type I, II, and or other)
Peripheral arterial disease (PAD)	Severe PAD (ICD codes including claudication, ulceration, gangrene, and rest pain)
Chronic obstructive pulmonary disease
Chronic liver disease	Severe chronic liver disease; at least one of cirrhosis and or hepatic failure any time prior to time of incident amputation
Heart failure
Myocardial infarction	In past 6 months
Peripheral neuropathy
Stroke	Without paraplegia, monoplegia, or paralysis
Vision loss
Retinopathy
Proteinuria
Microvascular disease	Requires diagnosis of peripheral neuropathy, retinopathy, and proteinuria
Mental health
Anxiety	Any anxiety disorder (including obsessive compulsive disorder, panic disorder, and or phobias)
Bipolar disorder	Any bipolar disorder (bipolar type I, II, and or other)
Depression	Ordinal category as no depression, depressive without major disorder, and major depressive disorder
Mild cognitive impairment
Psychosis	Includes schizophrenia and other psychotic disorders
Post-traumatic stress disorder
Health behaviours
Alcohol misuse in past year	Most recent Alcohol Use Disorders Identification Test—Consumption (AUDIT-C) score prior to time of incident amputation; categorised as mild, moderate, or severe
Smoking status	Ordinal category as never, former, and current
Opioid use disorder
Cocaine use disorder
Revascularisation§
Any ipsilateral revascularisation	Any ipsilateral revascularisation regardless of revascularisation type
Any contralateral revascularisation	Any contralateral revascularisation regardless of revascularisation type

BMI = body mass index; ICD = International Classification of Diseases.

^*A set of fractional polynomial terms was considered for all continuous predictors (age and BMI) to allow for non-linear association with (logit) risk.

^†Race and sex information were collected but were not included in the prediction model development because the number of eligible women was too small, and race may simply be a surrogate for other predictors.

^‡History of ever or never being diagnosed in the past, except for myocardial infarction.

^§Based on clinical experience that these procedures may influence healing and mobility.

Outcome

The seven item AMPSIMM was developed to classify a broad range of mobility in individuals following vascular amputation. It includes wheeled mobility and a specific distinction between home and different levels of community ambulation. It has demonstrated strong criterion and construct validity, and excellent responsiveness and floor and ceiling characteristics.²³ Similar to a previous publication,¹⁶ the seven category AMPSIMM was collapsed into a four category outcome to predict wheelchair mobility (0 – 2), and household (3), basic community (4), or advanced community ambulation (5 – 6).

Model development and validation

After data cleaning, the distributions of candidate predictors were evaluated to explore general patterns in the data. Categorical variables were tabulated, while continuous predictors were summarised with quartiles and measures of central tendency. Each predictor was also evaluated against the four category AMPSIMM scores in a bivariable manner.

One year post-amputation risk prediction models were fitted using multinomial logistic regression. Interactions of age (at time of amputation), marital status, BMI, chronic obstructive pulmonary disease, dialysis, depression, and any ipsilateral and contralateral revascularisations with all amputation levels were considered, as the level of amputation may induce differing effects on the risks of mobility.

Lasso regression was used for variable selection, which identifies more parsimonious models that optimise future predictive performance and that are tuned to reduce prediction error by minimising overfitting.²⁴ The optimal tuning parameter was determined using ten fold cross validation with a standard criterion that proposes the model minimising prediction error.²⁵ A more flexible “ungrouped” lasso model was chosen, allowing different sets of variables to predict each different category of AMPSIMM scores, if necessary. Since the outcome has multiple categories, individual coefficients do not represent log odds ratios as in binary logistic regression; their interpretability is more difficult, particularly for quantitative predictors, which are modelled non-linearly.

Due to the small amount of missingness (complete case observations n = 689, representing 2.1% of observations missing any predictor data), a complete case analysis was not conducted. For these observations, there were two predictors with missing data values: smoking status (2.1%) and Alcohol Use Disorders Identification Test – Consumption (AUDIT-C) status (0.4%). Multiple imputation by chained equations was implemented to handle these missing data. Both multiple imputation and lasso variable selection were performed simultaneously by stacking the imputed data sets and weighting observations by their proportions of non-missing variables.²⁶ Fractional polynomial terms were also included for the continuous predictors (age and BMI) to allow for a multitude of possible non-linear relationships with the outcome.²⁷

The final model was developed using the entirety of the data and was validated internally using optimism adjustment from 200 bootstrap samples. To assess the predictive ability of the model, a discrimination plot comprising groups of box plots that summarise predicted probabilities for each AMPSIMM category, stratified by observed outcome, was evaluated visually, and various quantitative performance measures were produced, including pairwise c indices for each pair of outcome categories, the average of these pairwise c indices (M index), a grouped c index for dichotomised AMPSIMM categories, and the polytomous discrimination index (PDI), which simultaneously compares sets rather than pairs of outcomes.²⁸ The PDI for random performance with a four category outcome is 0.25, whereas for c indices on category pairs it is 0.5. Statistical analyses were carried out using R statistical software with lasso multinomial regression performed using the glmnet package.^24,29

Model application

After model validation, the coefficients from the completed model were converted into a mobility calculator to demonstrate how the model could be used clinically. It is important to note that in the clinical application of the model, it predicts mobility outcome if the patient survives 1 year post-amputation. The two hypothetical patients were established based on the participating investigators’ clinical experience in the care of patients undergoing vascular amputation. Patient 1 is representative of a patient who requires amputation primarily secondary to complications of diabetes. This patient has type 2 diabetes, is obese, and has both renal failure and retinopathy. Patient 2, in contrast, requires amputation secondary to complications of PAD. He is lean, smokes, has had numerous revascularisations, and has coronary artery disease.

RESULTS

Sample summary

Among the consecutive 4 035 persons with an amputation, 1 790 (44%) were excluded (Fig. 1). This left 2 245 unilateral amputations eligible to be contacted for participation. Of 2 245 eligible participants, 240 became eligible after enrolment was closed and 128 had inadequate mailing addresses. Letters and questionnaires were posted to the 1 877 remaining eligible participants and phone calls were made to 1 630 eligible participants. From the mailed letters and questionnaires, 58 opted out and 189 completed the questionnaires sent initially. Of 1 630 phone calls, 427 completed the interview. The number of participants providing mobility outcome data was 704 (427 by phone and 277 by mailed questionnaire), a 37.5% recruitment rate. The distribution of amputation levels, outcomes, and candidate predictors are included in Table 2. Of note, those eligible who did not consent to participate were very similar to those who consented (Supplementary Table S2).

Figure 1.

Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) diagram showing the total number of veteran patients identified in the Veterans Affairs (VA) Corporate Data Warehouse (CDW), total numbers and reasons for exclusion, and final number enrolled in the study through mailings and phone interviews. * All consecutive transmetatarsal, transtibial, and transfemoral amputations due to peripheral arterial disease and or diabetes identified during the time period through the VA CDW. ^? Exclusions are in order they were applied, horizontally from left to right. ^‡ Patients with spinal cord injury codes that did not have a hemi- or quadriplegia code. ^§ Prior to their 12 month mobility outcome assessment. ^¶ Consented and enrolled, but no Amputee Single Item Mobility Measure (AMPSIMM) data collected.

Table 2.

Distribution of candidate predictors by Amputee Single Item Mobility Measure (AMPSIMM) category for eligible subjects (n = 704).

Variable	Wheelchair mobility (n = 217)*	Household mobility (n = 131)*	Basic community mobility (n = 196)*	Advanced community mobility (n = 160)*
Amputation level
Transmetatarsal	31 (14.3)	45 (34.4)	87 (44.4)	81 (50.6)
Transtibial	121 (55.8)	68 (51.9)	85 (43.4)	75 (46.9)
Transfemoral	65 (30.0)	18 (13.7)	24 (12.2)	4 (2.5)
Demographics
Age – y†	69.0 (62, 74)	67.0 (61, 73)	68.0 (61, 73)	64.0 (60, 74)
Sex‡
Male	212 (97.7)	128 (97.7)	195 (99.5)	158 (98.8)
Female	5 (2.3)	3 (2.3)	1 (0.5)	2 (1.2)
Race and ethnicity ‡
White, non-Hispanic	121 (58.5)	80 (63.5)	120 (65.2)	111 (73.5)
Black	67 (32.4)	40 (31.7)	45 (24.5)	27 (17.9)
Hispanic	12 (5.8)	3 (2.4)	12 (6.5)	10 (6.6)
Other	7 (3.4)	3 (2.4)	7 (3.8)	3 (2.0)
Marital status
Married	97 (44.7)	58 (44.3)	76 (38.8)	75 (46.9)
Single or never married	30 (13.8)	21 (16.0)	29 (14.8)	25 (15.6)
Separated or divorced	72 (33.2)	43 (32.8)	74 (37.8)	55 (34.4)
Widowed	18 (8.3)	9 (6.9)	17 (8.7)	5 (3.1)
Urban or rural classification
Rural	50 (23.0)	39 (29.8)	68 (34.7)	62 (38.8)
Urban	167 (77.0)	92 (70.2)	128 (65.3)	98 (61.3)
Geographical region
Continental	32 (14.7)	19 (14.5)	36 (18.4)	24 (15.0)
Midwest	45 (20.7)	27 (20.6)	51 (26.0)	39 (24.4)
North Atlantic	36 (16.6)	36 (27.5)	39 (19.9)	36 (22.5)
Pacific	41 (18.9)	20 (15.3)	41 (20.9)	32 (20.0)
Southeast	63 (29.0)	29 (22.1)	29 (14.8)	29 (18.1)
BMI — kg/m2†	28.1 (23, 32)	27.4 (24, 34)	27.4 (24, 31)	29.4 (25, 32)
Comorbidities §
Asthma	25 (11.5)	17 (13.0)	15 (7.7)	13 (8.1)
COPD	94 (43.3)	61 (46.6)	75 (38.3)	41 (25.6)
Kidney dialysis	30 (13.8)	9 (6.9)	15 (7.7)	7 (4.4)
Coronary atherosclerosis	130 (59.9)	73 (55.7)	93 (47.4)	75 (46.9)
Diabetes	188 (86.6)	118 (90.1)	172 (87.8)	152 (95.0)
Peripheral arterial disease	154 (71.0)	95 (72.5)	108 (55.1)	63 (39.4)
Chronic liver disease	13 (6.0)	15 (11.5)	12 (6.1)	6 (3.8)
Heart failure	101 (46.5)	59 (45.0)	68 (34.7)	39 (24.4)
Myocardial infarction	16 (7.4)	8 (6.1)	10 (5.1)	9 (5.6)
Peripheral neuropathy	142 (65.4)	89 (67.9)	129 (65.8)	102 (63.7)
Retinopathy	131 (60.4)	70 (53.4)	112 (57.1)	89 (55.6)
Proteinuria	45 (20.7)	24 (18.3)	34 (17.3)	18 (11.2)
Microvascular disease	35 (16.1)	20 (15.3)	26 (13.3)	14 (8.8)
Vision loss	29 (13.4)	9 (6.9)	20 (10.2)	10 (6.2)
Mental health §
Anxiety	85 (39.2)	52 (39.7)	77 (39.3)	49 (30.6)
Depression
No depression	97 (44.7)	54 (41.2)	97 (49.5)	86 (53.8)
Depressive disorder or dysthymia	34 (15.7)	22 (16.8)	40 (20.4)	30 (18.8)
Major depressive disorder	86 (39.6)	55 (42.0)	59 (30.1)	44 (27.5)
Any bipolar disorder	22 (10.1)	13 (9.9)	13 (6.6)	15 (9.4)
Psychosis or schizophrenia disorders	19 (8.8)	8 (6.1)	8 (4.1)	4 (2.5)
PTSD	62 (28.6)	32 (24.4)	44 (22.4)	33 (20.6)
Mild cognitive impairment	11 (5.1)	7 (5.3)	12 (6.1)	11 (6.9)
Health behaviours
AUDIT-C score‖
Mild	199 (92.1)	118 (90.8)	177 (90.8)	137 (85.6)
Moderate	13 (6.0)	11 (8.5)	15 (7.7)	20 (12.5)
Severe	4 (1.9)	1 (0.8)	3 (1.5)	3 (1.9)
Smoking status
Never	95 (44.6)	52 (40.6)	72 (37.5)	89 (57.1)
Former	60 (28.2)	29 (22.7)	65 (33.9)	33 (21.2)
Current	58 (27.2)	47 (36.7)	55 (28.6)	34 (21.8)
Cocaine use disorder	31 (14.3)	14 (10.7)	16 (8.2)	5 (3.1)
Opioid use disorder	14 (6.5)	12 (9.2)	21 (10.7)	6 (3.8)
Revascularisation
Any ipsilateral	97 (44.7)	58 (44.3)	81 (41.3)	39 (24.4)
Any contralateral	19 (8.8)	7 (5.3)	14 (7.1)	7 (4.4)

Data are presented as n (%, within column counts) or median (interquartile range). Median (interquartile range) was computed for all continuous predictors unless indicated otherwise. BMI = body mass index; COPD = chronic obstructive pulmonary disease; PTSD = post-traumatic stress disorder; AUDIT-C = Alcohol Use Disorders Identification Test — Consumption.

^*Wheelchair mobility: unable to walk but can use a wheelchair to get around your home and possibly in the community. Household mobility: able to walk inside your home with ambulation aids (e.g., cane, crutches, walker) but need a wheelchair to get around the community. Basic community mobility: able to walk limited distances in the community with ambulation aids (e.g., cane, crutches, walker). Advanced community mobility: able to walk in the community without the need for ambulation aids. This may range from a block to unlimited distances, like a shopping mall.

^†At the time of amputation.

^‡Race and sex are included in the table for generalisability but were not evaluated as candidate predictors in the model.

^§Unless otherwise specified, all diagnoses represent any diagnosis prior to the date of amputation.

^‖Most recent prior to the date of amputation (mild: women <3, men <4; moderate: women ≥3 and <8, men ≥4 and <8; severe >8 for both men and women). When multiple scores are recorded on the same date and time, the higher score is used.

Summary of outcomes distribution

Two hundred and seventeen individuals (30.8%) reported AMPSIMM scores indicating wheelchair mobility (score 0 – 2), 131 (18.6%) indicating household ambulation (score 3), 196 (27.8%) basic community ambulation (score 4), and 160 (22.7%) advanced community ambulation (score 5 – 6). Among TM amputees, this distribution was 12.7%, 18.4%, 35.7%, and 33.2%, respectively. Among TT amputees it was 34.7%, 19.5%, 24.4%, and 21.5%, and among TF amputees it was 58.6%, 16.5%, 21.6%, and 3.6%, respectively.

Risk prediction model development

The final prediction model included 15 predictors; estimated coefficients of predictors for each AMPSIMM mobility outcome category are provided in Table 3. Generally, the more positive the coefficients, the greater the associated probability for those specific outcome categories, and the more negative, the lower the associated probability. Continuous predictors modelled non-linearly (age and BMI) are illustrated graphically in Supplementary Figure S1.

Table 3.

Prediction coefficients for the final model of four category Amputee Single Item Mobility Measure (AMPSIMM) score outcome.

Variable	Wheelchair mobility (n = 217)	Household mobility (n = 131)	Basic community mobility (n = 196)	Advanced community mobility (n = 160)
Intercept	−0.003	−0.626	−0.075	0.704
Demographics
Age — y
x2 log x (×10−4)	0	0	0	−0.127
x2 (10 3)	0	0	0	−0.108
BMI — kg/m2
x2 (×10−4)	0	0	−0.972	0
x log x (×10−3)	0	0	−0.269	0
Comorbidities
COPD	0	0	0	−0.060
Kidney dialysis	0.035	0	0	0
Peripheral arterial disease	0.014	0.079	−0.014	−0.427
Chronic liver disease	0	0.035	0	0
Heart failure	0	0	0	−0.188
Retinopathy	0.002	0	0	0
Health behaviour
AUDIT-C score
Moderate or severe	0	0	0	0.044
Smoking status
Former or current	0	0	0.028	−0.093
Cocaine use disorder	0.115	0	0	0
Revascularisation
Any ipsilateral	0	0	0	−0.013
Interaction with TM amputation level
BMI
x log x	0	0	0	0.109
x−2	0	0	7.865	0
x−1 log x	−6.628	0	0	0
Interaction with TT amputation level
Kidney dialysis	0.450	0	0	0
Interaction with TF amputation level
Marital status, four category
Married	0.132	0	0	0
Widowed	0.068	0	0	0
Marital status, binary
Married	0.001	0	0	0
COPD	0.435	0	0	0
BMI — kg/m2
x log x (×10−2)	0.283	0	0	0
x(×10−3)	0.974	0	0	0
Age — y
x3 log x (×10−7)	0.123	0	0	0
x3 (×10−6)	0.143	0	0	0
x2 log x (×10−6)	0.734	0	0	0
x−1 log x	0	0	0	−1.422
x−1/2 log x	0	0	0	−0.014
x−1/2	0	0	0	−1.481

BMI = body mass index; COPD = chronic obstructive pulmonary disease; AUDIT-C = Alcohol Use Disorders Identification Test — Consumption; TM = transmetatarsal; TT = transtibial; TF = transfemoral; AMPSIMM = Amputee Single Item Mobility Measure.

The risk score, S_j (for j = 1,2,3,4), corresponding to an individual having a j category of AMPSIMM score outcome, is the sum of the coefficients in that category for all the components that apply to that individual (including a set of fractional polynomial terms for continuous predictors such as age). The predicted probability of j category of outcome = exp(S_j)/[exp(S₁) + exp(S₂) + exp(S₃) + exp(S₄)].

Model validation

Summaries of the various discrimination measures computed from the final model are shown in Table 4. Bootstrap estimates for optimism ranged 0.05 – 0.10 among the different measures, indicating some inflation due to overfitting from using the same data to develop and validate the model. The overall discrimination was relatively modest (M index = 0.66) after optimism adjustment, but pairwise c indexes indicated that the model is strongest at distinguishing individuals from each extreme of the outcome spectrum (i.e., wheelchair mobility from advanced community ambulation) (for AMPSIMM scores 0 – 2 vs.5 – 6, adjusted c index = 0.80). In contrast, the model was weaker at discriminating between household ambulation and basic community ambulation (for scores 3 vs. 4, adjusted c index = 0.55). However, the model was stronger at distinguishing household mobility or lower (scores 0 – 3) from community ambulators (4 – 6), and basic community ambulators or lower (0 – 4) from advanced community ambulators (5 – 6), with adjusted hierarchical c indices of 0.70 and 0.74, respectively. These are all important functional distinctions.

Table 4.

Quantitative discrimination measures for the final model.

Discrimination measure	Model estimate	Optimism adjusted
PDI	0.501	0.396
M index	0.731	0.655
Pairwise c index
Wheelchair vs. household mobility	0.682	0.594
Wheelchair vs. basic community mobility	0.745	0.679
Wheelchair vs. advanced community mobility	0.844	0.797
Household vs. basic community mobility	0.644	0.550
Household vs. advanced community mobility	0.767	0.688
Basic vs. advanced community mobility	0.703	0.623
Grouped c index
Wheelchair + household mobility vs. basic + advanced community mobility	0.752	0.700

PDI = polytomous discrimination index.

The discrimination plot for the final model is shown in Figure 2. The model is properly discriminating when the box plot of the observed outcome has the highest probability in the group corresponding to the predicted outcome category. The greatest separation in box plots (the greatest discriminative ability) is shown for individuals with wheelchair mobility (0 – 2) relative to advanced community ambulation (5 – 6), in accordance with having the highest the c index (Table 4). There is also expected separation in box plots for each of the mobility levels except for household vs. basic community mobility.

Figure 2.

Distributions of predicted probabilities for each Amputee Single Item Mobility Measure (AMPSIMM) category stratified by the actually observed category. The red lines show the prevalence of each AMPSIMM category.

Model application

The AMPREDICT Mobility-4 model is designed for use pre-amputation in a patient requiring LLA for complications of PAD and or diabetes. Assuming the patient survives 1 year after amputation, it can predict their probability of achieving four levels of functional mobility. Table 5 describes two representative patients and their clinical backgrounds pre-amputation. Patient 1 presents with a foot ulcer at the first metatarsophalangeal joint that extends to bone and is being considered for either a TM or TT amputation. The AMPREDICT Mobility-4 prediction model demonstrates that if the patient receives a TM amputation, they are most likely to be an advanced community ambulator (37%) and have at least a 63% chance of achieving some (when adding 37% to 26%) level of community mobility, contrasted with the 18% probability of being restricted to a wheelchair. Patient 2 is a 60 is 60 year old patient with PAD who is being considered for either a TT or TF amputation. The prediction reflects the well established adverse mobility impact of TF amputation compared with TT. The AMPREDICT Mobility-4 model suggests the patient has a 34% vs. 56% chance of being restricted to a wheelchair if they undergo TT amputation and TF amputation, respectively. At the TT level, this patient has a 47% chance of being at least a basic community ambulator (29% plus 18%) compared with 30% (21% plus 10%) if they undergo a TF amputation.

Table 5.

Example clinical scenarios.

Predictor or outcomes	Patient 1	Patient 2
Predictor
Age at time of amputation — y	70	60
Body mass index — kg/m2	38	22
Marital status	Single/never married	Married
Heart failure	No	Yes
Kidney dialysis	Yes	No
Diabetes	Yes	No
Peripheral arterial disease	No	Yes
Chronic liver disease	No	No
Chronic obstructive pulmonary disease	No	Yes
Retinopathy	Yes	No
Ipsilateral revascularisation	No	Yes
Alcohol misuse, AUDIT-C	Moderate	Moderate
Smoking status	Never	Current
Cocaine use disorder	No	No
Mobility outcomes
Wheelchair mobility — %
Transmetatarsal	18	16
Transtibial	42	34
Transfermoral	46	56
Household mobility — %
Transmetatarsal	18	24
Transtibial	14	19
Transfermoral	14	14
Basic community mobility — %
Transmetatarsal	26	36
Transtibial	20	29
Transfermoral	21	21
Advanced community mobility — %
Transmetatarsal	37	24
Transtibial	25	18
Transfermoral	19	10

AUDIT-C = Alcohol Use Disorders Identification Test — Consumption.

DISCUSSION

The primary purpose of this study was to develop and validate a patient specific multivariable model used at the time of amputation decision to predict a four category mobility outcome 1 year after amputation for individuals undergoing their first unilateral amputation at the TM, TT, or TF level secondary to diabetes and or PAD. The model uses a parsimonious set of 15 predictors that are readily available in the EHR. Model performance is modest for a four category outcome (M index = 0.71 and 0.64 with optimism adjustment); however, important functional level comparisons were stronger. It discriminated best when comparing wheelchair mobility vs. at least some community ambulatory mobility (optimism adjusted c index = 0.71), a basic level of community mobility vs. less than a basic level of community mobility (optimism adjusted c index = 0.70), and advanced community mobility vs. less than advanced community mobility (optimism adjusted c index = 0.73). This AMPREDICT Mobility-4 model is recommended to replace the original model, as (1) the data set contained nearly four times as many participants as the original, (2) all predictors are universally available within the EHR, thereby reducing clinician burden, and (3) the model predicts a more granular four category outcome, which provides more detailed information when considering rehabilitation and discharge planning.

A recent systematic review evaluating the accuracy of surgeon clinical outcome prediction in a variety of surgical conditions concluded that surgeons are generally good at predicting peri-operative risk, but poorer at predicting longer term outcomes. This review recommended the use of decision support tools to augment clinical decision-making.³⁰ A more recent publication reporting on the 1 year results of the PERCEIVE study confirmed this uncertainty, suggesting that prediction models may provide important support to shared decision-making around the amputation level decision.³¹ There are several prediction models published predicting future mobility in the vascular amputee population; however, these were developed post-amputation, at the time of discharge from rehabilitation and or at the time of prosthesis prescription.^31–35

McGinnis et al. recently published a study describing the benefits of a physician led collaborative care model for patients facing LLA, which involves a multidisciplinary team of surgeons, physiotherapists, prosthetists, and the patient, with the goal of enhancing amputation outcomes including mobility.³⁶ The authors noted that a collaborative care model serves to fulfil many best practice recommendations of the 2019 global vascular guidelines on the management of chronic limb threatening ischaemia.³⁷ The integration of well validated prediction models into this multidisciplinary approach can help to ensure that providers have the same personalised risk information for each patient. One of the global vascular guideline recommendations is for any non-urgent amputation to be discussed at a multidisciplinary team meeting after a full functional ability and vascular assessment. Having patient specific mobility predictions at each amputation level allows for more informed amputation level decision-making as well as post-surgical rehabilitation and prosthetic provision planning.

Another global vascular guideline states that patients should be informed as to the rationale of any amputation, as well as the post-amputation care pathway. It has been noted that many patients undergoing vascular surgery procedures³⁸ and amputation surgery^1,39 do not feel that they are offered the opportunity to participate in and desire greater involvement in this process.⁴⁰ Prediction models, particularly if converted into decision support tools, offer opportunities for patient engagement in shared decision-making around these important decisions.^41,42 A third guideline recommends pre-operative assessment by a rehabilitation and occupational physiotherapist, as well as by a prosthetic specialist. Here again, a clear patient specific understanding of probable mobility outcome will allow for better rehabilitation and prosthetic planning.^43,44

The original AMPREDICT-Mobility model⁴⁵ predicted two levels of mobility, defined by specific mobility tasks using the Locomotor Capabilities Index-5. This previous model predicted the equivalent of the current model’s basic community ambulation and advanced community mobility combined without differentiating between the two. Furthermore, it did not differentiate lesser mobility categories. The new AMPREDICT Mobility-4 model offers these important additions that more fundamentally address the patient concerns of “will I only be able to function from a wheelchair” or “will I be able to engage in at least household mobility”, while still allowing the prediction of basic community or advanced community ambulation. It allows for prediction of a greater spectrum of mobility outcomes than the previous model, providing greater granularity in setting patient expectations and rehabilitation planning. Furthermore, this new model does not require patient interview, as all predictors are readily available in the EHR.

The model has several strengths as well as potential limitations. Like the original AMPREDICT-Mobility model, this model predicts mobility 12 months after amputation and is only valid in patients who survive a full year. The model should be used alongside the AMPREDICT Mortality and Re-amputation model (AMPREDICT MoRe)¹⁵ so that providers and patients can balance the risk of death with achieving specific mobility levels should they survive. These models will be combined into a single calculator and incorporated into the online AMPREDICT Decision Support Tool (https://www.ampdecide.org). The number of women in the development model was very small because of its veteran population. However, it was decided a priori not to model sex or race to avoid inappropriate application of the model that could contribute to disparities in decision-making should these factors influence key outcomes. Since the model was designed to inform decisions prior to amputation, events that occurred after amputation that may affect future mobility (e.g., re-amputation, contralateral amputation, or other significant comorbid events) could not be accounted for. Due to a moderate sample size, internal validation of the model was performed to preserve as much data as possible for model training and increased predictive power. Using a bootstrap method with many predictors may result in an underestimation of optimism, but the simulations suggested that this underestimation should be minimal. Finally, the primary outcome was self reported mobility as opposed to a direct assessment of patient mobility. Clinician evaluation of mobility in the household and community is not realistic and would be of high clinical burden, and self reported mobility is common and accepted in many patient populations. Like the previously published AMPREDICT models,^4,16,19,45 this model carefully ensured that the amputations were first amputations and did not include re-amputations or contralateral amputations. Furthermore, predictors were carefully selected to include those readily available in the EHR to reduce the clinical burden. Although there was a 38% response rate, the characteristics of patients who were eligible but not enrolled were very similar to those enrolled, reducing potential concerns of response bias. Finally, while this was a national sample of US veterans from many geographical regions, the uniqueness of the veteran population may limit some of the model’s generalisability. It is recommended that the model be externally validated in other populations.

The next steps in the development will be to convert this model into a decision support tool. The AMPREDICT Decision Support Tool,⁴⁶ along with the AMPDECIDE patient decision aids,^47,48 are currently being used internationally to provide education and evidence for informing these decisions and facilitating shared decision-making (https://www.ampdecide.org). The AMPREDICT MoRe models have recently been redeveloped and published using a more contemporary population and including predictors that are readily available in the EHR to facilitate clinical implementation of the model.¹⁵

Summary

This study has presented the development and validation of a novel AMPREDICT Mobility-4 model that predicts four categories of functional mobility to be applied at the time of amputation level decision-making using predictors available in the EHR. The model has been converted into a point of care decision support tool, available both externally for all users and within the VA EHR. This tool serves to augment providers’ clinical expertise in facilitating patient surgeon amputation level shared decision-making, developing rehabilitation plans, and assisting in setting appropriate patient outcome expectations.

WHAT THIS PAPER ADDS.

This paper references a new AMPREDICT-Mobility prediction model that predicts four categories of functional mobility among individuals facing their first amputation as a consequence of diabetes and or peripheral arterial disease. This model overcomes prior implementation barriers using only predictors readily available in the electronic health record. The model can augment clinical personalised mobility outcome prediction, enhancing both surgeon and patient awareness of the consequences of amputation level on mobility, to facilitate patient surgeon amputation level shared decision-making and assisting in setting appropriate rehabilitation goals and patient outcome expectations.

APPENDIX A. SUPPLEMENTARY DATA

Supplementary data to this article can be found online at https://doi.org/10.1016/j.ejvs.2025.06.041.